Why Applicant Tracking Systems Matter

There are two stories about applicant tracking systems. In the first, a robot reads your resume, fails to find the right keywords, and deletes you before a human ever knows you applied. In the second, ATS software is a harmless filing cabinet and worrying about it is a waste of time.

Both are wrong, and the gap between them is where the useful information lives.

The software is effectively universal

Jobscan's 2025 usage report checked the career pages of every company on the Fortune 500 list as of its June 2, 2025 release and detected an applicant tracking system at 489 of them — 97.8%. The eleven exceptions most likely run something proprietary rather than nothing at all. Workday alone accounts for 39% of that group; across a wider sample of 12,820 companies the market is more fragmented, led by Greenhouse at 19.3% and Lever at 16.6%.

So the question is not whether your application passes through a system. It is what the system is asked to do with it.

What employers say their systems do

The most direct evidence comes from the employers themselves. Harvard Business School's Project on Managing the Future of Work, working with Accenture, surveyed employers and workers for its 2021 report Hidden Workers: Untapped Talent. More than 90% of employers surveyed said they use their recruiting management system to filter or rank candidates for middle-skills and high-skills roles.

Then comes the finding that matters:

The report attributes this to how the criteria are written and configured — a degree requirement on a role that never needed one, a hard filter on continuous employment, an exact-match job title — rather than to a parser deciding you are unqualified. The authors estimate the aggregate effect at roughly 27 million "hidden workers" in the United States: people who want to work and are looking, but whose applications keep dying against filters that were never really about their ability to do the job.

That is the honest version of the fear. It is not that a machine reads your resume badly. It is that a human wrote a rule, and the machine applies it without judgment.

The 75% statistic is not real

You have seen the claim that ATS software rejects 75% of resumes before a human sees them. It appears in career advice, in vendor marketing, and in a great many LinkedIn posts.

It traces back to Preptel, a resume-optimization company, in around 2012. No methodology was ever published. Preptel went out of business the following year. Every subsequent citation is a citation of a citation, and the chain terminates at a company that was selling the cure for the disease it described.

Discard the 75% figure. The HBS survey numbers above are better, they come with a named methodology, and they point somewhere more actionable.

What actually changed: the software started judging

For most of their history, applicant tracking systems sorted and searched. The recruiter still decided. That is no longer a safe assumption, and the research on what happens when models do the deciding is not reassuring.

Kyra Wilson and Aylin Caliskan of the University of Washington tested three production large language models on résumé screening, presenting the work at the 2024 AAAI/ACM Conference on AI, Ethics, and Society. They varied 120 first names associated with white and Black men and women across 554 real resumes and more than 500 real job listings, generating over three million resume-to-job comparisons. The models favored white-associated names 85% of the time against 9% for Black-associated names, and male-associated names 52% of the time against 11% for female-associated names. They never preferred a Black male-associated name over a white male-associated one.

This is not a claim that your specific application was scored by a biased model. It is a claim about what the tooling does when it is pointed at resumes without supervision — and it is the reason the regulators arrived.

Regulators now treat screening software as a selection procedure

• New York City Local Law 144 requires employers using an automated employment decision tool to commission an independent bias audit annually and publish the results. (NYC rules)
• The EEOC issued guidance in May 2023 confirming that algorithmic decision-making tools are "selection procedures" under Title VII and the Uniform Guidelines on Employee Selection Procedures — meaning disparate-impact liability attaches to the employer, not the vendor. That guidance was removed from the agency's website in January 2025 following an executive order. It was a technical-assistance document explaining existing law rather than creating it, so Title VII still applies; only the explanation is gone. (archived copy)
• Illinois HB 3773, signed August 2024 and effective January 1, 2026, amended the Illinois Human Rights Act to prohibit employers from using AI that has a discriminatory effect in recruitment and hiring, and to require notice when AI is used in employment decisions. (Illinois General Assembly)
• Colorado enacted the first comprehensive state AI law covering consequential decisions including employment, though its obligations have since been narrowed and its effective date pushed back more than once. (Littler analysis)

A jurisdiction does not require annual bias audits of a filing cabinet. The regulatory attention is itself evidence about what these systems are doing — and with federal guidance withdrawn, the rules that bind an employer now depend substantially on which state the job is in.

What follows from all this

The practical conclusions are narrower than the anxiety suggests.

Your resume needs to survive the parse. That is a mechanical problem with a mechanical fix, and it is covered in detail in how applicant tracking systems actually read your resume.

Your resume needs to match the posting's stated criteria where those criteria are actually true of you. The 88% finding is not about keyword density. It is about the specific, checkable things a filter can be configured on: job titles, years of experience, credentials, tools named in the requirements. If you have the thing and the posting names it differently than you do, use their words.

Do not optimize against the machine. Hidden keywords in white text are extracted along with everything else, and the mismatch between extracted and visible text is itself detectable. Meanwhile the thing that determines the outcome — whether a person reading your application believes you can do the job — is unaffected by any of it.

The system is real, it is nearly universal, and its influence is narrower and more boring than the folklore claims. Write for the human, but make sure the software can find you first.

Sources

• Jobscan, 2025 Applicant Tracking System Usage Report (Fortune 500 career-page audit, June 2025).
• Joseph B. Fuller and Manjari Raman et al., Harvard Business School and Accenture, Hidden Workers: Untapped Talent, September 2021.
• Kyra Wilson and Aylin Caliskan, "Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval," AAAI/ACM Conference on AI, Ethics, and Society, 2024 — summarized by UW News.
• New York City, Local Law 144 / Automated Employment Decision Tools rules.
• U.S. Equal Employment Opportunity Commission, "Select Issues: Assessing Adverse Impact in Software, Algorithms, and Artificial Intelligence Used in Employment Selection Procedures Under Title VII," May 18, 2023 — removed from eeoc.gov in January 2025; archived copy.
• Illinois General Assembly, HB 3773, 103rd General Assembly.